ATBECHDEDKESFRGBGRITLILUNLSEMCPTIE......FI....CY................................JDIM350 (Ver 2.1 Jan 2001) 2100000/00977173EUROPEAN PATENT SPECIFICATIONB120030521EP99202506.4199907302001090320011119enenen94856 P19980731US2003052120032120000202200005200305212003217 7G 10L 15/08 AdeMinimierung eines Suchnetzwerks für die SpracherkennungenMinimization of search network in speech recognitionfrMinimisation d'un réseau de recherche pour la reconnaissance de la paroleEP-B- 0 248 377DESHMUKH N ET AL: "Efficient search strategies in hierarchical pattern recognition systems" PROCEEDINGS OF THE ANNUAL SOUTHEASTERN SYMPOSIUM ON SYSTEM THEORY, 1995, XP002126397HIRAYAMA H ET AL: "DSP-BASED LARGE VOCABULARY SPEAKER-INDEPENDENT SPEECH RECOGNITION" NEC RESEARCH AND DEVELOPMENT,JP,NIPPON ELECTRIC LTD. TOKYO, vol. 37, no. 4, 1 October 1996 (1996-10-01), pages 528-534, XP000685766 ISSN: 0547-051XHANAZAWA K ET AL: "AN EFFICIENT SEARCH METHOD FOR LARGE-VOCABULARY CONTINUOUS-SPEECH RECOGNITION" IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING (ICASSP),US,NEW YORK, IEEE, 21 April 1997 (1997-04-21), pages 1787-1790, XP000735007 ISBN: 0-8186-7920-45Yo-Hong, Kao1515 Rio Grande No. 1609Plano, Texas 75075USTexas Instruments Incorporated02688370TI-27219 EU7839 Churchill WayDallas, Texas 75251USHolt, Michael00050421Texas Instruments Limited, P.O. Box 5069Northampton NN4 7ZEGBATBECHCYDEDKESFIFRGBGRIEITLILUMCNLPTSE20010314200111 TECHNICAL FIELD OF THE INVENTION

This invention relates to speech recognition and more particularly to minimization of a search network.

BACKGROUND OF THE INVENTION

Continuous speech recognition is a resource-intensive algorithm. Commercial dictation software requires more than 10M bytes to install on the disk and 32M bytes RAM to run the application. Static memory is required to store the program (algorithm), grammar, dictionary, and acoustic models. The data will not change, therefore can be stored in disk or ROM. Dynamic memory is required to run the search. Search involves parsing the input speech and building a dynamically changing search tree, therefore, RAM is required for both Read and Write capabilities. The target application for this patent disclosure is for small vocabulary speech recognition implemented on, for example, fixed-point DSP chips. On a DSP implementation, RAM size is a critical cost factor because RAM occupies a large percentage of the total die size. ROM, on the other hand, is much cheaper in silicon real estate.

EP 0 248 377 B1 discloses a continuous speech recognition system in which frames of input speech are matched against templates. Each time a potential match for a word is found a node representing the word is added to a tree, alternative possible matches for some frames appearing in the tree as alternative branches. Each node also records a measure of the similarity between the sequence of input frames and the templates leading up to that node. A traceback algorithm prunes the tree of nonuseful information to prevent information from excessively accumulating in memory; in particular it removes records that are no longer under consideration as a potential match.

SUMMARY OF THE INVENTION

According to the invention, there are provided a method as set out in claim 1 and a speech recognizer as set out in claim 5.

In accordance with one embodiment of the present invention, in the search network expansion in speech recognition only the current slot is maintained as long as we know what model and what state the slot is at and the previous slots can be discarded.

DESCRIPTION OF THE DRAWINGS

DESCRIPTION OF PREFERRED EMBODIMENTS OF THE PRESENT INVENTION

Referring to Fig. 1 there is illustrated a speech recognition system. The recognizer 10 includes a library 10a including a pronunciation dictionary, grammars, acoustic models, etc. The recognizer 10 further includes a computer/comparator 10b for comparing parsed input speech to the models and for computing a score and memory 10c for storing operation program and storing results from comparing and computing results. The parsed speech input is compared to speech models and the match produces the recognized output. The framework of our recognition system is HMM (Hidden Markov Model). Where the sentence grammar is represented by a Markov model with states 12 and transitions 11 among states (see Fig. 2). Transitions are associated with words. When a transition is taken from state A to state B, one of the words associated with that transition must be evaluated. Then from state B, there are again many outgoing transitions to choose from, each transition has words associated with it. Taking a transition means going through a word. This Markov model thus describes what words a sentence can start with, what words follow what words, and what words a sentence can end with. This is a computational representation of a grammar.

Each word is also represented by a Markov model with states and transitions. Acoustics are associated with each state. Transition into a state means evaluating the acoustics associated with that state. Usually a left-to-right HMM is used for word model, there are straight-through transitions 11 representing average speed of speaking, self-loop transitions 13 representing slow speed, and skip transitions 17 representing fast speed. Acoustics can also be associated with transitions (as in sentence HMM); however, in most speech recognition systems the acoustics are associated with states for its simplicity.

These two levels of HMM's describe the search space of a speech recognition system (see Y.H. Kao, W. Anderson, H.S. Lim, "A Multi-Lingual, Speaker-Independent, Continuous-Speech Recognizer on TMS320C5x Fixed-Point DSP," ICSPAT 1997, San Diego, USA and Y.H. Kao, "Fixed-Point Implementation of IG Speech Recognizer on C5x DSP," TI Tech Report, 1996). From the top level sentence grammar to the bottom level acoustics, the recognition algorithm (parser) can run the input acoustic vectors through this search space to find out the best path by building a search network. The best path found at the end of the input vectors is the recognition result. Grammars can be represented by context-free-grammar (for small vocabulary tasks) or N-Gram (for large vocabulary tasks). In a large vocabulary system, usually a three level system (sentence, word, phone) is used rather than the two level system (sentence, word). It is impossible to build individual word models for so many words, therefore, phonetic models are used as the base units (see Y.H. Kao, K. Kondo, "Phonetic Modeling Using Acoustic Decision Tree," TI Tech Report, 1997; Y.H. Kao, "Acoustic Decision Tree: A Tutorial," TI Tech Report, 1997; Y.H. Kao, "Acoustic Decision Tree: Performance Analysis," TI Tech Report, 1997).

Search is expanding all the possible paths in the grammar (see Fig. 3). When a speech frame comes in, first we expand all the possible word in the sentence HMM. To expand each word, we need to expand their individual phone sequence in the lexicon HMM. To expand each phone, we need to expand their phonetic HMM's, which has acoustics as the observations. There are three levels of HMM in the structure. The upper level transitions may take more than one speech frame, only the bottom level transitions consume exactly one speech frame. A speech frame is, for example, 20 milliseconds long. The upper level transition can be taken only after its corresponding observation is completed (which may take several speech frames).

Speech recognition is to expand a search network following the grammar-word search space definition. There are many ways of actually defining the computer data structure and implementing the algorithm. We will use our algorithm as an example and describe the ways to minimize the network.

We define a structure called slot as the building block of the search network. Using a C structure, it is defined as:

Model_index is an integer representing what model this slot is at. For example, the whole search space is represented by one sentence model, we assign model_index 0 to the sentence model. This sentence model is composed of many words and how these words can be put together, we assign model_index 1, 2, 3, ... etc. for each of the word models. Model_index can not overlap, each index represents a different model (sentence or words).

State_index is an integer representing what state (in a model, be it sentence or words) this slot is at. Both sentence and word models are HMM's, they are evaluated state by state. When building the search network, we need to know which state we are at, so we know what next state(s) to transition to. Within each model, state_index starts from 1, 2, 3, ... etc. State 1 in model 1 is different from state 1 in model 2.

Score is the accumulated score of this path up to this slot. Backptr is a pointer that points back to the previous slot in the path. For example, if state 10 can come from state 9 (straight-through transition), or state 8 (skip transition), or state 10 (self-loop transition); after Viterbi decoding which keeps only the best path coming into a state, the backptr of the state 10 slot will point to one of the three slots described above.

Time is the time index when this slot is first created. For example, we use 20 ms frame length. The input speech is segmented into 20 ms frames, preprocessed to extract a feature vector, then fed into the search algorithm. During the search of the first frame (0-20 ms), time is 1; during the search of the second frame (20-40 ms), time is 2, it goes on like this.

Last_time is the last time this path is updated. We need this time stamp for slot management (garbage collection). Expansion of the search network is an exponential growth problem, bad scored paths have to be pruned to reduce the size of search network. When a path has good score and should be retained for future expansion, we propagate the current time stamp backward through the whole path (a path is a backward linked list of slots). If the last_time of a slot is the current time, it must be retained and can not be reused. Otherwise, it can be reused because its path is not within the search beam and thus last_time not updated.

Next_state points to the next slot of active state within this model to be evaluated. When evaluating a model, many states can be active and need to be evaluated, they are linked together by next_state.

Next_word points to the next slot of active word for this sentence state to be evaluated. When evaluating a sentence model, its slots of active states are linked by next_state. But for each state, it has words that are still being evaluated (a word needs more than one frame to complete). Next_word links all these pending word evaluation slots. Next_word will start from a sentence level slot.

Search is the most complicated part of the speech recognition algorithm. The best way to learn the algorithm is to trace the C code. Please refer to our well commented C code and a companion documentation (see Y.H. Kao, "IG (Integrated Grammar) Algorithm," TI Tech Report, 1996).

Fig. 4 illustrates an example of the search space for the sentence grammar "Call George Washington," The layers of grammars are sentences, words, phones and acoustic distributions (bottom layer) represented by the small circles. The expansion goes from word "call" to phone "|K|" to the top three circles 41-43 for the acoustics of "|K|". The expansion then goes to the second phone "|ao|" to the five circles 44 and then back to the phone "|l|". The expansion then goes to the acoustics for "|l|" represented by the small circles 45. After the last circle 45, the expansion goes to the word "George" then to the phone "|jh|". The expansion continues with the three acoustics 47, then the phone "|ao|" and then the five acoustics 49. After the last acoustic 49, the search goes to the phone "|r|", followed by the four acoustics followed by the phone "|jh|", followed by the three acoustics 53. The last acoustic 53 is followed by the word "Washington" followed by the phone "|w|". This is followed by the three acoustics 55. The acoustics 55 are followed by the phone "|ao|" followed by the five acoustics 57, followed by the phone "|sh|". The phone "|sh|" is followed by its four acoustics 59, followed by the phone "|ax|" and its three acoustics 61. The expansion continues similarly for the phones "|n|", "|t|", "|ax|", and "|n|" followed by their associated three acoustics.

For the sake of continuing our discussion, we will describe the idea of search in a high level, and then explain the search network minimization process.

After defining the slot data structure as the building block of the search network, it is straightforward to do the search network expansion. It can be summarized as:

For each input acoustic vector (20 ms), we have to traverse down the search space as shown, for example, in Fig. 4 to reach the bottom layer so an acoustic evaluation can be done. The acoustic score and the transition score are accumulated and stored in the score field of the slot (put in storage 10c of Fig. 1). The backptr field points to previous slot (where this slot comes from). Therefore, for each input acoustic vector, the evaluation will add at least one slot to the search path because one acoustic evaluation has to be done. Sometimes more than one slot has to be added to the path because sentence level slot has to be added before the acoustic evaluation slot can be added.

This is called trace mode, every input acoustic vector is mapped into a model acoustic vector and recorded in the search path. It is very expensive because a very long linked list of slots is created. For example, a 5-second input utterance has 5 x 50 = 250 input vectors. A linked list of more than 250 slots has to be created for every possible theory within the search beam width.

For training purpose, this is necessary because every input vector needs to be mapped to a model vector in order to update that model vector. But for recognition purpose, this is an overkill because we only want to know what words are recognized. We do not need to know the microscopic trail of how each vector is mapped.

Now let's study the search network expansion in more detail; specifically, we would like to examine the reason for keeping slots in the process.

For the purpose of expansion, only the current slot needs to be maintained in storage 10c (Fig. 1). A slot provides the model_index and the state_index in that model. As long as we know what model (sentence or words) and what state this slot is at, we can consult the model and figure out what next states to transition to. There is no need to know the previous slots for the search network expansion purpose. By this argument, only one slot, the current slot, has to be maintained in a path; all previous slots in a path can be discarded.

However, for the recognition purpose, we want to know what words are recognized after the expansion is completed. Because of this reason, one slot has to be maintained for each recognized word in a search path. For the example of Fig. 4, after "call" has been recognized, a slot for the word "call" is all that is maintained. All the other expansion of phones and acoustics can be discarded. A control signal between computer/comparator 10b and storage 10c controls the discard. This is done after computation. Compute all slots but discard previous slot since present slot has accumulated score and once a word is determined we only need the word. The back pointer points to previous word.

By the above argument, for a 10-digit recognition, we only have to keep 10 slots in the search path. Each slot represents one recognized digit. This is the shortest possible path we can do to solve the 10-digit search problem.

Referring to Fig. 5 there is illustrated a flow chart for expansion and keeping and discarding slots. The input speech is expanded in step 101 and compared to the library models and dictionary 10a at computer/comparator 10b at step 103 and the score is placed in the slots and the score is accumulated. The present slot scores are kept and poor score paths are discarded. When a phone end is detected at step 105 it is determined if a word end. If a word end is not detected, the previous slots are discarded and the score accumulated in the present slot. When a word is detected (step 107), a word slot is created (step 109) for the identified word and all previous slots are discarded and the identified word is built into the search network. The back pointer then points to the previous word.

Unfortunately, the one-word-one-slot only network does not carry any sentence level information. Sentence level information is necessary to represent grammar, which describes how words can be put together. At the minimum, one sentence level slot at the word boundary is required to provide the information to carry on the search into the next word. Without this sentence level slot, we do not know what next word to evaluate when we complete evaluating the current word. In addition to carrying the grammar information, sentence level slot is also used to carry N-Best information and provide a base for sentence level Viterbi decoding.

Armed with the above understanding of the search mechanism, we can now bound the length of a search path given a grammar. For example, for a single word recognition grammar, the shortest possible path consists of one word, three slots:

        S1WS2

where W is the word and S1 is the silence before the word and S2 the silence after the word W.

The longest possible path is 4N + 3 where there are N words (not counting silence), N + 1 silence slots and 2N + 2 sentence slots. The longest path of three words (beginning silence + the recognized word + ending silence) is seven slots:

        S1WsilS1WS2WsilS2

where W represents a word level slot, S represents a silence level slot.

For a 10-digit grammar, the shortest possible path has 10 digits + 11 sentence slots = 21 slots. The longest possible path has 10 digits + 11 silence + 22 sentence slots = 43 slots. It is independent of the length of the input utterance, it only depends on the grammar.

This minimization principle can be easily generalized into the three level (sentence, word, phone) search structure for large vocabulary task.

In continuous speech recognition, silence is considered a word. Grammar allows optional silence word in all word boundaries. However, the between word acoustics can vary a lot - silence or background noises. Usually we use a large number of mixtures in the silence model in order to capture the widely varying acoustics.

To minimize the search path, we want to let one silence model capture all the between word acoustics. We do not want to allow several different silence models to appear in one between word segment because it will lengthen the path unnecessarily.

We calibrated the peak slot usage on two tasks: single word military alphabet recognition (7087 files) and 10 continuous digits recognition (1390 files).

Two conditions were run for these two tasks:

Peak slot usage histograms are plotted in Fig. 6. In the plots, they are represented by circle and plus respectively. The X-axis is the peak slot count, the Y-axis is the number of utterances that requires minimum X-axis slots to be successfully parsed. The more left the distribution is, the more efficient the search is.

For the single word military alphabet recognition task, the max, mean, and standard deviation of the peak slot usage are:

For the 10 continuous digits recognition task, the max, mean, and standard deviation of the peak slot usage are: The optimally minimized search network reduces about 20% of the peak slot usage from our previous effort for both tasks.

We proposed ways to minimize the search network within the framework of our search algorithm. The search path length can be bounded by the grammar and is independent of input utterance length. It achieved 20% reduction in search space compared to our previous non-optimal effort. Compared to the trace mode, the reduction is 70% to 80%. Usually the trace mode is done under supervision, therefore, the peak slot usage will not be nearly as large. However, if unsupervised model update is needed, the peak slot usage will be very high, as depicted in the histogram plots.

In this application we used a slot structure with 8 fields. In another application, we will describe a way to reduce that into 5 fields. Combining the search network minimization described in this application and the slot size reduction, the 10-digit task can be done with only 1600 slots x 5 words = 8K words of search space (RAM) (see Y.H. Kao, "Reducing Slot Size in Speech Recognition Search Network by Data Packing and a New Slot Management Mechanism," TI Tech Report, 1998). In our 1390 10-digit utterances, only two utterances require more than 1600 slots to parse.

 A method of providing a search network in speech recognition comprising the steps of: parsing input speech; building a dynamically changing search tree by expanding all active search paths and creating time dated slots containing model index, model state, time index indicating the time the slot was created, a last time index indicating the last time the slot was updated and an accumulated score of matching to speech models of the path up to the slot and comparing-the parsed input speech to speech models and identifying what model and what model state and computing said accumulated score of matching; discarding bad score paths and bad scored slots; propagating a current time stamp backward through a linked list of slots for paths with good scores to provide a last time index with the current time at said slots in said paths with good scores; and discarding slots and paths not having said current time at said last time index. The method of Claim 1 including the step of determining the end of a word and creating a word end slot when a word end is determined and discarding all previous slots. The method of Claim 1 or Claim 2, wherein each slot further stores the following: a back pointer, next state and next word. The method of Claim 1, Claim 2 or Claim 3, including the step of limiting the longest possible path to 4N + 3 "word. level or silence level slots" where there are N words (not counting silence) at most. N+1 silence words plus (2N + 1) +1 sentence slots. A speech recognizer for input speech comprising: a storage; a library of grammar and speech models; a computer/comparator coupled to said library and responsive to said input speech for building a search tree by expanding active search paths and creating dated time slots containing model index, model state, time index indicating the time the slot was created, a last time index indicating the time the slot was last updated and an accumulated score of matching to speech models of the path up to the slot, said computer/comparator computing a score for said match and an accumulated score at each slot and said computer/comparator coupled to said storage for storing said model index, model state, time index, last time index and said accumulated score of matching; discarding bad score paths and bad scored slots; propogating a current time stamp backward through a linked list of slots for paths with good scores to provide a last time index with the current time at said slots in said paths with good scores; and discarding slots and paths not having said current time at said last time index. The speech recognizer of Claim 5, wherein said computer/comparator determines the end of a word and creates a word slot and discards all previous slots. The speech recognizer of Claim 5 or 6, wherein for each slot there is further stored a model index, a back pointer, a next state and a next word. The speech recognizer of Claim 5, Claim 6 or Claim 7, wherein said computer/comparator permits only one silence slot between words. Verfahren zur Schaffung eines Suchnetzes in der Spracherkennung, das die folgenden Schritte umfaßt: Parsen der Eingangssprache; Aufbauen eines sich dynamisch ändernden Suchbaums durch Erweitern aller aktiven Suchpfade und Erzeugen von zeitgestempelten Schlitzen, die einen Modellindex, einen Modellzustand, einen Zeitindex, der den Zeitpunkt angibt, zu dem der Schlitz erzeugt wurde, wobei ein letzter Zeitindex den letzten Zeitpunkt angibt, zu dem der Schlitz aktualisiert wurde, und einen akkumulierten Übereinstimmungsstand mit Sprachmodellen des Pfades bis zu dem Schlitz enthalten, Vergleichen der geparsten Eingangssprache mit Sprachmodellen und Identifizieren, welches Modell und welcher Modellzustand in Frage kommen, und Berechnen des akkumulierten Übereinstimmungsstandes; Verwerfen von Pfaden mit schlechtem Ergebnis und von Schlitzen mit schlechtem Stand; Ausbreiten eines aktuellen Zeitstempels rückwärts durch eine verknüpfte Liste von Schlitzen für Pfade mit guten Ständen, um bei den Schlitzen in den Pfaden mit guten Ständen einen letzten Zeitindex mit der aktuellen Zeit bereitzustellen; und Verwerfen von Schlitzen und Pfaden, die nicht diese aktuelle Zeit als den letzten Zeitindex haben. Verfahren nach Anspruch 1, das den Schritt umfaßt, bei dem das Ende eines Wortes bestimmt wird und ein Wortendeschlitz erzeugt wird, wenn ein Wortende bestimmt wird, und alle vorhergehenden Schlitze verworfen werden. Verfahren nach Anspruch 1 oder Anspruch 2, bei dem jeder Schlitz ferner einen Rückwärtszeiger, einen nächsten Zustand und ein nächstes Wort speichert. Verfahren nach Anspruch 1, Anspruch 2 oder Anspruch 3, das den Schritt umfaßt, bei dem der längstmögliche Pfad auf 4N + 3 "Wortpegel- oder Ruhepegel-Schlitze" begrenzt ist, wenn N Wörter (ohne Zählung der Ruhe), höchstens N + 1 Ruhewörter plus (2N+1) + 1-Satzschlitze vorhanden sind. Spracherkenner für Eingangssprache, der umfaßt: einen Speicher; eine Grammatik- und Sprachmodell-Bibliothek; einen Computer/Komparator, der mit der Bibliothek gekoppelt ist und auf die Eingangssprache reagiert, um einen Suchbaum durch Erweitern aktiver Suchpfade aufzubauen und zeitgestempelte Schlitze zu erzeugen, die einen Modellindex, einen Modellzustand, einen Zeitindex, der den Zeitpunkt angibt, zu dem der Schlitz erzeugt wurde, wobei ein letzter Zeitindex den Zeitpunkt angibt, zu dem der Schlitz zuletzt aktualisiert wurde, und einen akkumulierten Übereinstimmungsstand mit Sprachmodellen des Pfades bis zu dem Schlitz enthalten, wobei der Computer/Komparator einen Übereinstimmungsstand und einen akkumulierten Stand bei jedem Schlitz berechnet und wobei der Computer/Komparator mit dem Speicher gekoppelt ist, um den Modellindex, den Modellzustand, den Zeitindex, den letzten Zeitindex und den akkumulierten Übereinstimmungsstand zu speichern; wobei Pfade mit schlechtem Stand und Schlitze mit schlechtem Stand verworfen werden;    wobei ein aktueller Zeitstempel nach rückwärts durch eine verknüpfte Liste von Schlitzen für Pfade mit guten Ständen ausgebreitet wird, um bei den Schlitzen in den Pfaden mit guten Ständen einen letzten Zeitindex mit der aktuellen Zeit bereitzustellen; und
   wobei Schlitze und Pfade, die nicht die aktuelle Zeit als den letzten Zeitindex haben, verworfen werden. Spracherkenner nach Anspruch 5, bei dem der Computer/Komparator das Ende eines Wortes bestimmt und einen Wortschlitz erzeugt und alle vorhergehenden Schlitze verwirft. Spracherkenner nach Anspruch 5 oder 6, bei dem für jeden Schlitz ferner ein Modellindex, ein Rückwärtszeiger, ein nächster Zustand und ein nächstes Wort gespeichert werden. Spracherkenner nach Anspruch 5, Anspruch 6 oder Anspruch 7, bei dem der Computer/Komparator nur einen Ruheschlitz zwischen Wörtern zuläßt. Procédé de création d'un réseau de recherche en reconnaissance de parole, comprenant les étapes consistant à : analyser grammaticalement la parole d'entrée ; construire un arbre de recherche variant dynamiquement en étendant tous les trajets actifs de recherche et en créant des intervalles datés de temps contenant un index de modèle, un état de modèle, un index de temps indiquant l'instant où l'intervalle a été créé, un index de dernier instant indiquant le dernier instant où l'intervalle a été actualisé, et un score cumulé de concordance avec des modèles de parole du trajet jusqu'à l'intervalle, et en comparant la parole analysée d'entrée à des modèles de parole et identifiant quel modèle et quel état de modèle, et calculant ledit score cumulé de concordance ; éliminer des trajets de mauvais score et les intervalles de mauvais score ; propager un horodatage courant en remontant dans une liste d'intervalles munie de liens pour des trajets ayant de bons scores afin d'obtenir un index de dernier instant avec l'instant actuel auxdits intervalles dans lesdits trajets ayant de bons scores ; et éliminer des intervalles et des trajets n'ayant pas ledit instant courant audit index de dernier instant. Procédé selon la revendication 1, comprenant l'étape de détermination de la fin d'un mot et de création d'un intervalle de fin de mot quand une fin de mot est déterminée, et d'élimination de tous les intervalles précédents. Procédé selon la revendication 1 ou la revendication 2, dans lequel chaque intervalle stocke en outre ce qui suit : un pointeur arrière, un état suivant et un mot suivant. Procédé selon la revendication 1, la revendication 2 ou la revendication 3, comprenant l'étape de limitation du trajet le plus long possible à 4N + 3 fois des intervalles de valeur de mot ou de valeur de silence, où il existe N mots (sans compter le silence) à au plus N + 1 mots de silence plus (2N + 1) + 1 intervalles de phrase. Dispositif de reconnaissance de parole pour une parole d'entrée, comprenant : une mémoire ; une bibliothèque de modèles de grammaire et de parole ; un calculateur-comparateur couplé à ladite bibliothèque et répondant à ladite parole d'entrée pour construire un arbre de recherche en étendant des trajets actifs de recherche et en créant des intervalles datés de temps contenant un index de modèle, un état de modèle, un index de temps indiquant l'instant où l'intervalle a été créé, un index de dernier instant indiquant le dernier instant où l'intervalle a été actualisé, et un score cumulé de concordance avec des modèles de parole du trajet jusqu'à l'intervalle, ledit calculateur-comparateur calculant un score pour ladite concordance et un score cumulé à chaque intervalle et ledit calculateur-comparateur étant couplé à ladite mémoire pour stocker lesdits index de modèle, état de modèle, index de temps, index de dernier instant et ledit score cumulé de concordance ; éliminant les trajets de mauvais score et les intervalles de mauvais score ; propager en arrière un horodatage de temps courant dans une liste munie de liens d'intervalles pour des trajets ayant un bon score afin d'obtenir un index de dernier instant avec l'instant actuel auxdits intervalles dans lesdits trajets ayant de bons scores ; et éliminer des intervalles et des trajets n'ayant pas ledit instant courant audit index de dernier instant. Dispositif de reconnaissance de parole selon la revendication 5, dans lequel ledit calculateur-comparateur détermine la fin d'un mot et crée un intervalle de mot et élimine tous les intervalles précédents. Dispositif de reconnaissance de parole selon la revendication 5 ou la revendication 6, dans lequel pour chaque intervalle il est stocké en outre un index de modèle, un pointeur arrière, un état suivant et un mot suivant. Dispositif de reconnaissance de parole selon la revendication 5, la revendication 6 ou la revendication 7, dans lequel ledit calculateur-comparateur autorise seulement un intervalle de silence entre des mots.